Always compare the model to the current process (the baseline), at the action's capacity.
Generic top-K value model
K = cases acted on per period (capacity)
prec@K = share of acted cases that are true positives
v_tp = value of acting on a true positive (net of action effect: saves x value)
c_fp = cost of acting on a false positive
c_action = cost of the action itself (call, review)
value(model) = K * (prec@K * v_tp - (1 - prec@K) * c_fp - c_action)
lift = value(model) - value(baseline)
ceiling = value at prec@K = min(1, positives_per_period / K)
break-even prec@K = (c_action + c_fp + run_cost/K) / (v_tp + c_fp)Churn retention
| Input | Value |
|---|---|
| Capacity | 400 calls/week |
| Churners per week | 300 (out of 50,000 active) |
| Save rate when a churner is called | 30% |
| Customer value if saved | $600 |
| Call cost | $8 |
| Baseline prec@400 | 0.25 |
| Model prec@400 | 0.40 |
Saves: baseline 400 x 0.25 x 0.30 = 30/week; model 48/week. Lift 18 x $600 = $10,800/week. Call cost is the same in both. Ceiling: prec@400 = 300/400 = 0.75 -> 90 saves/week, so a perfect model adds at most $36k/week over baseline.
Caveat: if calling also annoys non-churners (c_fp > 0) or the save rate differs by segment, use uplift: the value comes from persuadable customers, not from churners in general.
Fraud at authorisation
| Input | Value |
|---|---|
| Transactions/day | 2,000,000 |
| Fraud rate | 0.1% (2,000/day), average $150 |
| Decline cost (false positive) | lost margin $3 + friction; customer churn risk |
| Review capacity | 2,000 cases/day |
Report fraud dollars caught at a fixed false-decline rate (e.g. 0.2% of good transactions), not AUC.
Value = fraud $ prevented - (false declines x cost per false decline) - review cost. Because fraud amounts
vary, rank by expected loss (p_fraud x amount) and measure value-weighted recall.
Demand forecasting
Value comes from inventory: lower safety stock for the same service level, fewer stock-outs, less waste.
safety_stock ~ z_service * sigma_forecast_error * sqrt(lead_time)
holding_saving = (sigma_old - sigma_new) * z * sqrt(L) * unit_cost * holding_rateA 10% reduction in forecast error standard deviation reduces safety stock by ~10%. Measure error at the decision's aggregation (store-SKU-week), weighted by margin or volume.
Lead scoring
value = (opportunities from top-K leads under model - under baseline order) x win rate x deal valueMeasure on leads that were actually worked; beware that reps already prioritise (selection bias): run an A/B on queue order.
Cost of errors table
Write this table with the business owner; it becomes the threshold choice later.
| Outcome | Business effect | Cost estimate |
|---|---|---|
| True positive | action prevents loss | +$ value saved |
| False positive | action wasted, possible harm | -$ action cost - friction |
| False negative | loss not prevented | -$ loss (relative to perfect) |
| True negative | nothing | 0 |
The optimal threshold for a calibrated probability p is to act when p * v_tp > (1 - p) * c_fp + c_action,
i.e. p > (c_fp + c_action) / (v_tp + c_fp); with a capacity, take the top K instead.
Build and run cost
| Item | Typical size |
|---|---|
| Data engineering for features and labels | often 50-70% of the effort |
| Modelling and validation | 10-20% |
| Serving, monitoring, retraining | 20-30%, recurring |
| Human review / action cost | per prediction, recurring |
A project is feasible when the expected lift over its first year exceeds build plus run cost with margin for the uncertainty in the lift estimate (use the lower end of the metric CI).
anomaly-detection-model-builder
Build and evaluate anomaly detection for tabular, sensor and transaction data with few or no labels: define the anomaly type (point, contextual, collective), choose detectors (robust z-score/MAD and IQR rules, Mahalanobis with robust covariance, k-NN distance, LOF, Isolation Forest, one-class SVM, PCA reconstruction / autoencoders, forecasting residuals for time series), combine them by rank, explain each alert by its deviating features, set thresholds from an alert budget or the score distribution, and evaluate with precision@k, PR-AUC and recall on known incidents, or detector agreement when no labels exist. Includes pure-Python detectors (robust z, Mahalanobis, kNN, Isolation Forest, ensemble) and an evaluation script.
catboost-model-builder
Build CatBoost models that use its strengths correctly: raw categorical columns as cat_features (ordered target statistics, no manual encoding), text_features, Pool construction with NaN-safe categories, ordered boosting and has_time for temporal data, symmetric (oblivious) trees, overfitting detector with eval_set and use_best_model, losses for classification, regression (Quantile, Poisson, Tweedie) and ranking (YetiRank), class weights, monotone constraints, feature importance (PredictionValuesChange, LossFunctionChange) and SHAP values, GPU training and model export. Plans cat/text/numeric columns from a CSV and generates a runnable training script, and checks parameter sets for the conflicts CatBoost rejects (iterations vs n_estimators, Bayesian bootstrap with subsample, multiple class-weight options, depth > 16, use_best_model without eval_set).
categorical-encoding-selector
Choose and implement the right encoding for every categorical column given the model family and the data: one-hot (with infrequent-level grouping and handle_unknown), ordinal with an explicit order, cross-fitted target encoding (scikit-learn TargetEncoder, CatBoost ordered statistics, category_encoders inside CV), frequency/count encoding, feature hashing, native categorical support in LightGBM, HistGradientBoosting, XGBoost and CatBoost, prefix levels for codes (zip3, NAICS-2), and embeddings. Measures cardinality, rare-level share, unseen-level rate on a time-ordered holdout and target signal per column, emits the ColumnTransformer, and quantifies target-encoding leakage (in-sample vs out-of-fold vs holdout AUC).
classification-model-selector
Choose the right classification algorithm for tabular, text or mixed data and prove the choice with a fair bake-off: logistic regression (L2, L1, elastic net), linear SVM, naive Bayes, k-NN, RBF SVM, decision trees, random forest, extra trees, HistGradientBoosting, LightGBM, XGBoost, CatBoost, explainable boosting machines (EBM/GAM), points scorecards and MLPs, matched to data size, width, categorical and high-cardinality share, missingness, class imbalance, text, and constraints such as interpretability, monotonic constraints, calibrated probabilities, latency budgets and tiny deployable models. Ranks candidates with starting hyperparameters and cautions, then generates a bake-off script running every candidate on identical CV folds with per-fold metrics, fit time and prediction latency.
clustering-algorithm-selector
Choose the right clustering algorithm and check that clustering makes sense at all: cluster tendency (Hopkins statistic with a correlated-features caveat, nearest-neighbour isolation), then k-means / MiniBatchKMeans, Gaussian mixtures with BIC, agglomerative (Ward, average, complete, single) with dendrograms, DBSCAN, HDBSCAN, OPTICS, spectral clustering, BIRCH, mean shift, k-prototypes and k-modes for mixed or categorical data, and Gower distance, matched to data size, expected shapes, noise, varying density, known k, soft memberships, hierarchy and streaming. Covers preprocessing (scaling, encoding, dimensionality, distance metrics), validation (internal indices, bootstrap stability, external labels, business profiling) and assigning new points.
collaborative-filtering-implementer
Implement collaborative filtering recommenders that beat popularity: item-based kNN with cosine, Jaccard or asymmetric similarity, shrinkage, neighbourhood size, positive-only or mean-centred rating signals and 'because you liked X' explanations; implicit ALS with confidence weighting, BPR pairwise ranking and the closed-form EASE model in plain NumPy, tuned on a temporal validation split by recall@k and NDCG@k and refit on all data; plus user-kNN, explicit matrix factorisation, the implicit/LightFM libraries, hyperparameter ranges, cold-start fallbacks, scaling with sparse matrices and approximate nearest neighbours. Produces top-k lists in the format of the recommendation evaluator.